Anthropic Challenges Amazon with its Claude Agents, but Is It Enough?

Anthropic Launches Claude Managed Agents: A Bold Technological Advancement

On April 8, Anthropic unveiled the beta version of Claude Managed Agents, a new platform that promises to revolutionize the management of intelligent agents. This launch has been widely covered by the media, highlighting notable improvements such as a tenfold increase in execution speed. Companies like Notion and Asana have already adopted this technology, which offers controlled sessions and sandbox execution. Anthropic has also introduced a tool call management system, simplifying tasks for developers.

The technical article accompanying this launch highlights a major innovation: the separation of the agent stack into stable abstractions, reminiscent of hardware virtualization by operating systems in the 1990s. This approach allows for more flexible and scalable management of agents.

The results of this innovation are significant: the wait time for the first token has been reduced by approximately 60%, and the p95, which measures the performance of the slowest 5% of cases, has improved by over 90%. These figures attest to the increased efficiency of the system.

Key Innovations from Anthropic

Anthropic has designed Managed Agents as a hosted execution environment that is both robust and intuitive. Users can define an agent by specifying its system prompt, tools, and guardrails using simple formats like YAML or natural language. Sessions are designed to persist over several days, and tool calls occur in secure environments. Identifiers are protected in vaults, invisible to the sandbox, thereby ensuring enhanced security.

The pricing model is usage-based, with a cost of $0.08 per hour of active session, in addition to the standard rates for Claude tokens. Companies like Notion use this technology to delegate tasks to Claude within their workspaces, while Rakuten has developed agents for various departments, integrated via Slack and Teams. Sentry, on the other hand, pairs its debugging agent with a Claude agent to automate bug fixes.

An interesting aspect of this architecture is how sandboxes are managed. Viewed as cattle rather than pets, they are provisioned on demand, allowing for increased flexibility and scalability.

The Importance of Anthropic's Architecture

Anthropic's architecture stands out with its session model as an event log, an innovation that allows for state storage outside the harness. This approach ensures session continuity, even in the event of a crash, thanks to the wake(sessionId) command. This prevents the model's context window from becoming the primary storage, thereby reducing the risks of silent and costly errors.

This solution is particularly relevant for those who have previously lost agents due to context overflows. The isolation of identifiers is another crucial detail at the production scale. By grouping identifiers in the sandbox at the time of provisioning, Anthropic avoids potential errors where an agent might access sensitive information it should not see.

Challenges of Competing with Amazon

However, Anthropic's launch comes five months after a similar advancement from Amazon. Amazon Bedrock AgentCore reached general availability in late 2025, and its SDK has been downloaded over two million times in five months. Each AgentCore session operates in an isolated microVM, with sessions lasting up to eight hours. The execution environment is flexible, capable of hosting various frameworks like LangGraph or CrewAI.

A Defensive Strategy for Anthropic

In this context, Anthropic's launch appears as a defensive response to competition. Amazon's AgentCore can already host a Claude agent, just like Vertex. Anthropic must therefore convince its clients to remain loyal to its model, despite the possibility of AWS offering more competitive rates for session hours.

Anthropic's strategy seems aimed at strengthening its developer base, preventing them from migrating to competing solutions. While the company focuses on selling Claude tokens, the execution layer it offers is more of a distribution method than a true breakthrough innovation.

The Analogy with Operating Systems: A Long-Term Vision

Anthropic's technical article draws on a comparison with operating systems, suggesting that the separation of sessions, harnesses, and sandboxes into stable interfaces allows for independent evolution of each layer. This approach is reminiscent of hardware virtualization, but it also raises questions about the sustainability of this model in a market where virtualization has become a commodity.

Historically, VMware dominated the virtualization market for a decade before the technology became accessible to all. Anthropic may face a similar challenge, where its technology, although innovative, could quickly be caught up by more affordable competing solutions.